The first API built for Dictation: AssemblyAI's Dictation API
Read the Portuguese translation →
Author: AssemblyAI · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: The first API built for Dictation: AssemblyAI’s Dictation API · Channel: AssemblyAI · 03:16 · Watch on YouTube
Topics: Prompt and context engineering, AI developer tools, Multimodal models
Summary
Section titled “Summary”AssemblyAI presenta su Dictation API, diseñada para convertir audio en texto práctico y depurado, en lugar de limitarse a transcribir cada sonido literalmente. La demostración muestra resultados originales y limpios en distintos contextos, y destaca la velocidad, la consistencia y la precisión. La API funciona con Universal 3.5 Pro, admite más idiomas además del inglés y permite personalizar instrucciones y pasar contexto o términos clave para mejorar el reconocimiento.
Key points
Section titled “Key points”- La API de dictado está diseñada para generar texto que las personas puedan usar, no solo transcripciones literales.
- Admite 19 idiomas.
- Devuelve tanto el resultado original como una versión depurada, para que el producto elija cuál mostrar.
- La velocidad y la consistencia son aspectos centrales de la experiencia de dictado.
- Funciona con el modelo Universal 3.5 Pro, también utilizado en las API asíncronas y en tiempo real de AssemblyAI.
- La demostración incluye ejemplos de reuniones, notas de incidencias, recordatorios, listas y notas médicas.
- Se pueden personalizar las instrucciones para modificar cómo se limpia el resultado.
- El contexto de la conversación y los términos clave esperados pueden pasarse como indicaciones para mejorar el reconocimiento.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | El problema del dictado y la nueva API | AssemblyAI explica por qué la transcripción literal no basta para el dictado y presenta una API que devuelve texto utilizable en 19 idiomas. |
| 00:31 | Demostración con distintos usos | La demostración compara el audio con los resultados originales y depurados mientras muestra ejemplos de trabajo, recordatorios y listas. |
| 01:38 | Velocidad, consistencia e idiomas | El video subraya la importancia de la rapidez y la consistencia, y señala que la API se basa en Universal 3.5 Pro y admite otros idiomas además del inglés. |
| 02:44 | Personalización y contexto | AssemblyAI explica cómo personalizar la limpieza del texto y usar indicaciones, contexto y términos clave para mejorar el reconocimiento. |
Quotes
Section titled “Quotes”They are trained to capture every sound that comes out of your mouth. — 00:00
The dictation API gives you the original as well as the cleaned up results — 00:31
The first is speed. The whole point of dictation is to type three x faster than a keyboard. — 01:38
Out of the box, it cleans up the mess, but if you want something different, you can customize the instructions as well. — 02:44
You can pass context about the conversation or the key terms you expect — 03:16
Concepts and entities
Section titled “Concepts and entities”Dictation API, AssemblyAI, Universal 3.5 Pro, ditado, transcrição, prompt
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 You’ve probably seen that little microphone icon at this point on every app you use. That’s dictation showing up everywhere right now. And everyone building it hits the same wall, which is transcription models are verbatim. They are trained to capture every sound that comes out of your mouth. That’s okay for a transcript, but not okay for dictation. So that’s why we built our first dictation API. You send audio, you get back text that you’ll actually use in 19 [music] languages and shape the way your product needs it.
00:31 Let’s take a look. Um so so I think for the retry logic we should we should back off exponentially and cap it at 30 [music] seconds. Here’s what I said and here’s what my app gets. The dictation API gives you the original [music] as well as the cleaned up results and you decide which one your product shows. And the number on the corner right here is not [music] a marketing figure. It’s actually how long it took to give you the response. Now let me take a few more takes on different domains on purpose. Can you make our one-on-one move it to [music] Thursday afternoon?
01:03 Anything after two works. Note for the ticket, the crash only happens after cold start, so just fix it only for the iOS. Remind me to send Priya the updated pricing deck right before [music] Friday. We need almond milk, coffee filter, and good bread if they have it.
01:38 There are two things. The first is speed. The whole point of dictation is to type three x faster than a keyboard. Second is consistency. The user needs to have same [music] experience every single time. Okay, so fix [music] is in Lip Sodium. Can you can you ping Anjali on the SDK team. A key part of dictation is capturing words accurately. So, this runs on our flagship Universal 3.5 Pro model. The same one behind our async and real-time
02:11 APIs. This also means the dictation API supports more than just English. I mean the water try going to achieve 30 second may and failure go lock going to achieve. And it holds up on how people actually speak. >> Remind me to ask about my offer letter right before my 4:00. >> And that’s because we want it to work no matter where you are on a noisy street or in an office. Um so, 45-year-old guy, lower back pain,
02:44 um going down the left leg. It’s been about 2 weeks. We’re going to uh put him on a physio twice a week and see him in a month. Out of the box, it cleans up the mess, but if you want something different, you can customize the instructions as well. >> [music] >> So, the instruction cleans up the output you receive, but prompting shapes how the model hears it. Dictation runs on Universal 3.5 Pro. So, the prompt runs the same way like anywhere else on our API. [music]
03:16 You can pass context about the conversation or the key terms you expect, and the recognition improves before the cleanup ever runs. With the right prompt, you can see the right terms and entities [music] captured correctly. So, that’s it. With one call, you can quickly build dictation to any [music] product fast, customizable, and reliable. Our dictation API is available today, [music] and we can’t wait to see what you build with it. >> [music]